AI that survives production.

The problem

Analysts were reading thousands of sponsorship contracts by hand. Naive automation was not an option: a wrong extraction propagates into downstream reporting, and a job that dies late in the chain cannot afford to re-burn every prior stage of tokens.

The approach

Each document is a work unit with an explicit status enum and one background job per stage. A stage only enqueues the next on success, and every job is idempotent, so a failure resumes exactly where it broke. All model calls route through a provider gateway with bounded retries, a hard timeout and typed handling for read and connect timeouts. Before anything reaches reporting, the pipeline writes a QA record listing the model's disagreements and open issues and halts at a review gate — only that item blocks; the queue keeps draining.

The problem

A general-purpose model gives confident, generic coaching advice. The client's value was in their own methodology, so answers had to be grounded in their material and traceable back to it.

The approach

Source content is embedded into a vector index and retrieved per query, with the assistant orchestrated across multiple model providers so no single one is a point of failure. Content syncs from the client's existing operational tooling, so their team keeps editing where they already work. A scheduler drives recurring prompts without a human in the loop.

The problem

Latency is the whole product. A roleplay that pauses between turns stops being practice and starts being a form.

The approach

The audio round trip is split into independently tunable stages — transcription, model turn, synthesis — so each can be optimised or swapped without touching the others. Scenario state is carried across turns so the AI counterpart stays in character for the length of a real call.

The problem

Ad platform APIs are unstable, rate-limited and awkward to type. Letting a consumer app talk to one directly means every outage becomes a frontend bug.

The approach

A dedicated service sits in front of the ad platform, owning campaign, creative and metrics access behind a typed contract. The product app consumes that, streams model responses token by token so the interface never sits blank, and generates imagery through a separate inference provider.

The problem

Per-token cost is real and variable. Flat-rate pricing on a generative product is a way to lose money on your heaviest users.

The approach

Consumption is metered in the domain unit the customer understands — words — and credits are decremented against a single balance regardless of whether they arrived from a subscription or a one-off purchase. Document processing runs in the background with progress surfaced live.

The problem

Vision inference does not belong inside a Rails request cycle, and the Python ML ecosystem does not belong inside a Rails deployment.

The approach

The detection workload was split into a separate containerised service with a narrow HTTP contract. The Rails app stays the system of record for sessions and results; the vision service stays independently deployable and independently scalable.